Fault sample generation under continuous degradation using CRGAN-MDTR: A conditional recurrent GAN with maximum degenerate trend retention

Optimal data conditions are crucial prerequisites for the effective implementation of data-driven fault-diagnosis methods. However, the lack of sufficient training samples, especially under conditions of unknown fault severity levels, significantly hampers their performance in practical engineering applications. A generative adversarial network (GAN) offers a feasible solution, but existing GAN-based models face challenges in generating missing samples with varying fault degrees during continuous degradation. To address this issue, we propose a novel data-generation method, called Conditional Recurrent GAN with Maximum Degenerate Trend Retention (CRGAN-MDTR). The CRGAN integrates long short-term memory (LSTM) units into a Conditional GAN (CGAN) architecture, enhancing its ability to learn temporal dependencies from continuous degradation trends. In addition, a degradation trend reconstruction network (MDTR) is incorporated alongside the CRGAN discriminator, enabling controlled generation of fault samples and further improving their quality. The proposed CRGAN-MDTR was comprehensively evaluated through ablation studies and comparisons with commonly used generative techniques. The evaluation employed three categories of metrics: Sample Consistency, Fault Severity Consistency, and Classification Task-Oriented Metrics, to ensure robustness across multiple dimensions. Results demonstrated that CRGAN-MDTR consistently achieved superior performance, including improved consistency and diagnostic accuracy, with the highest accuracy increase reaching 4.7 %. Notably, the model maintained high robustness in noise response experiments and effectively interpolated fault samples at varying degradation levels, addressing a critical gap in fault diagnosis methodologies. These findings validate the robustness and effectiveness of CRGAN-MDTR in addressing fault data scarcity across varying fault severity levels. This study contributes a promising solution to enhancing fault diagnosis performance under extreme data scarcity conditions and provides a foundation for future research on real-world industrial applications.